The seasonal-trend disentangle based prognostic framework for PEM fuel cells

Abstract

Prognostics and Health Management (PHM) offer a potential solution to improve the durability of Proton Exchange Membrane Fuel Cells (PEMFCs), where degradation prognostic is the key component of the PHM system. However, limited by the dynamic operating environment, accurate and efficient prognostic algorithms still required for real applications. In this paper, the Seasonal-Trend Disentangle (STD) based prognostic framework is proposed to enhance the Remaining Useful Life (RUL) prediction of PEMFC system. Specifically, a lightweight parameter estimation neural network is introduced to estimate the aging parameters of the FC degradation model directly from PEMFC measurements. Besides, the integral of the FC degradation model at the specified current density range serves as the comprehensive Health Index (HI), which could combine different aging parameters and mitigate the effect of load dynamics. The STD-based prognostic model effectively captures the time-invariant features in the HI sequence from trend and seasonal phases, improving Remaining Useful Life (RUL) prediction performance. The proposed prognostic model is compared with RNN-family, ESN, and Dilated CNN methods using three distinct durability datasets from FCs operating under static, quasi-dynamic, and cyclically varying dynamic current loads. Experimental results demonstrate that the proposed algorithm outperforms the reference methods in terms of RMSE and MAPE metrics in the aging prognostic, and the proposed method could obtain satisfactory RUL prediction results 100–150 h earlier than the reference methods.